Environment Creation#

This documentation overviews creating new environments and relevant useful wrappers, utilities and tests included in Gym designed for the creation of new environments. You can clone gym-examples to play with the code that are presented here. We recommend that you use a virtual environment:

git clone https://github.com/Farama-Foundation/gym-examples
cd gym-examples
python -m venv .env
source .env/bin/activate
pip install -e .

Subclassing gym.Env#

Before learning how to create your own environment you should check out the documentation of Gym’s API.

We will be concerned with a subset of gym-examples that looks like this:

gym-examples/
  README.md
  setup.py
  gym_examples/
    __init__.py
    envs/
      __init__.py
      grid_world.py
    wrappers/
      __init__.py
      relative_position.py

To illustrate the process of subclassing gym.Env, we will implement a very simplistic game, called GridWorldEnv. We will write the code for our custom environment in gym-examples/gym_examples/envs/grid_world.py. The environment consists of a 2-dimensional square grid of fixed size (specified via the size parameter during construction). The agent can move vertically or horizontally between grid cells in each timestep. The goal of the agent is to navigate to a target on the grid that has been placed randomly at the beginning of the episode.

  • Observations provide the location of the target and agent.

  • There are 4 actions in our environment, corresponding to the movements “right”, “up”, “left”, and “down”.

  • A done signal is issued as soon as the agent has navigated to the grid cell where the target is located.

  • Rewards are binary and sparse, meaning that the immediate reward is always zero, unless the agent has reached the target, then it is 1.

An episode in this environment (with size=5) might look like this:

where the blue dot is the agent and the red square represents the target.

Let us look at the source code of GridWorldEnv piece by piece:

Declaration and Initialization#

Our custom environment will inherit from the abstract class gym.Env. You shouldn’t forget to add the metadata attribute to you class. There, you should specify the render-modes that are supported by your environment (e.g. "human", "rgb_array", "ansi") and the framerate at which your environment should be rendered. Every environment should supportNone as render-mode; you don’t need to add it in the metadata. In GridWorldEnv, we will support the modes “rgb_array” and “human” and render at 4 FPS.

The __init__ method of our environment will accept the integer size, that determines the size of the square grid. We will set up some variables for rendering and define self.observation_space and self.action_space. In our case, observations should provide information about the location of the agent and target on the 2-dimensional grid. We will choose to represent observations in the form of a dictionaries with keys "agent" and "target". An observation may look like {"agent": array([1, 0]), "target": array([0, 3])}. Since we have 4 actions in our environment (“right”, “up”, “left”, “down”), we will use Discrete(4) as an action space. Here is the declaration of GridWorldEnv and the implementation of __init__:

import gym
from gym import spaces
import numpy as np
from typing import Optional
from gym.utils.renderer import Renderer

class GridWorldEnv(gym.Env):
    metadata = {"render_modes": ["human", "rgb_array"], "render_fps": 4}

    def __init__(self, render_mode: Optional[str] = None, size: int = 5):
        assert render_mode is None or render_mode in self.metadata["render_modes"]

        self.size = size  # The size of the square grid
        self.window_size = 512  # The size of the PyGame window

        # Observations are dictionaries with the agent's and the target's location.
        # Each location is encoded as an element of {0, ..., `size`}^2, i.e. MultiDiscrete([size, size]).
        self.observation_space = spaces.Dict(
            {
                "agent": spaces.Box(0, size - 1, shape=(2,), dtype=int),
                "target": spaces.Box(0, size - 1, shape=(2,), dtype=int),
            }
        )

        # We have 4 actions, corresponding to "right", "up", "left", "down"
        self.action_space = spaces.Discrete(4)

        """
        The following dictionary maps abstract actions from `self.action_space` to 
        the direction we will walk in if that action is taken.
        I.e. 0 corresponds to "right", 1 to "up" etc.
        """
        self._action_to_direction = {
            0: np.array([1, 0]),
            1: np.array([0, 1]),
            2: np.array([-1, 0]),
            3: np.array([0, -1]),
        }

        """
        If human-rendering is used, `self.window` will be a reference
        to the window that we draw to. `self.clock` will be a clock that is used
        to ensure that the environment is rendered at the correct framerate in
        human-mode.
        """
        if render_mode == "human":
            import pygame  # import here to avoid pygame dependency with no render

            pygame.init()
            pygame.display.init()
            self.window = pygame.display.set_mode((self.window_size, self.window_size))
            self.clock = pygame.time.Clock()
                
        # The following line uses the util class Renderer to gather a collection of frames 
        # using a method that computes a single frame. We will define _render_frame below.
        self.renderer = Renderer(render_mode, self._render_frame)
            

Constructing Observations From Environment States#

Since we will need to compute observations both in reset and step, it is often convenient to have a (private) method _get_obs that translates the environment’s state into an observation. However, this is not mandatory and you may as well compute observations in reset and step separately:

    def _get_obs(self):
        return {"agent": self._agent_location, "target": self._target_location}

We can also implement a similar method for the auxiliary information that is returned by step and reset. In our case, we would like to provide the manhattan distance between the agent and the target:

    def _get_info(self):
        return {"distance": np.linalg.norm(self._agent_location - self._target_location, ord=1)}

Oftentimes, info will also contain some data that is only available inside the step method (e.g. individual reward terms). In that case, we would have to update the dictionary that is returned by _get_info in step.

Reset#

The reset method will be called to initiate a new episode. You may assume that the step method will not be called before reset has been called. Moreover, reset should be called whenever a done signal has been issued. Users may pass the seed keyword to reset to initialize any random number generator that is used by the environment to a deterministic state. It is recommended to use the random number generator self.np_random that is provided by the environment’s base class, gym.Env. If you only use this RNG, you do not need to worry much about seeding, but you need to remember to call super().reset(seed=seed) to make sure that gym.Env correctly seeds the RNG. Once this is done, we can randomly set the state of our environment. In our case, we randomly choose the agent’s location and the randomly sample target positions, until it does not coincide with the agent’s position.

The reset method should either return an observation of the initial state, or a tuple of the initial observation and some auxiliary information, depending on whether return_info was True. We can use the methods _get_obs and _get_info that we implemented earlier for that:

    def reset(self, seed=None, return_info=False, options=None):
        # We need the following line to seed self.np_random
        super().reset(seed=seed)

        # Choose the agent's location uniformly at random
        self._agent_location = self.np_random.integers(0, self.size, size=2)

        # We will sample the target's location randomly until it does not coincide with the agent's location
        self._target_location = self._agent_location
        while np.array_equal(self._target_location, self._agent_location):
            self._target_location = self.np_random.integers(0, self.size, size=2)

        # clean the render collection and add the initial frame
        self.renderer.reset()
        self.renderer.render_step()

        observation = self._get_obs()
        info = self._get_info()
        return (observation, info) if return_info else observation

Step#

The step method usually contains most of the logic of your environment. It accepts an action, computes the state of the environment after applying that action and returns the 4-tuple (observation, reward, done, info). Once the new state of the environment has been computed, we can check whether it is a terminal state and we set done accordingly. Since we are using sparse binary rewards in GridWorldEnv, computing reward is trivial once we know done. To gather observation and info, we can again make use of _get_obs and _get_info:

    def step(self, action):
        # Map the action (element of {0,1,2,3}) to the direction we walk in
        direction = self._action_to_direction[action]
        # We use `np.clip` to make sure we don't leave the grid
        self._agent_location = np.clip(
            self._agent_location + direction, 0, self.size - 1
        )
        # An episode is done if the agent has reached the target
        done = np.array_equal(self._agent_location, self._target_location)
        reward = 1 if done else 0  # Binary sparse rewards
        observation = self._get_obs()
        info = self._get_info()

        # add a frame to the render collection
        self.renderer.render_step()

        return observation, reward, done, info

Rendering#

Here, we are using PyGame for rendering. A similar approach to rendering is used in many environments that are included with Gym and you can use it as a skeleton for your own environments:

    def render(self):
        return self.renderer.get_renders()

    def _render_frame(self, mode: str):
        import pygame # avoid global pygame dependency. This method is not called with no-render.
    
        canvas = pygame.Surface((self.window_size, self.window_size))
        canvas.fill((255, 255, 255))
        pix_square_size = (
            self.window_size / self.size
        )  # The size of a single grid square in pixels

        # First we draw the target
        pygame.draw.rect(
            canvas,
            (255, 0, 0),
            pygame.Rect(
                pix_square_size * self._target_location,
                (pix_square_size, pix_square_size),
            ),
        )
        # Now we draw the agent
        pygame.draw.circle(
            canvas,
            (0, 0, 255),
            (self._agent_location + 0.5) * pix_square_size,
            pix_square_size / 3,
        )

        # Finally, add some gridlines
        for x in range(self.size + 1):
            pygame.draw.line(
                canvas,
                0,
                (0, pix_square_size * x),
                (self.window_size, pix_square_size * x),
                width=3,
            )
            pygame.draw.line(
                canvas,
                0,
                (pix_square_size * x, 0),
                (pix_square_size * x, self.window_size),
                width=3,
            )

        if mode == "human":
            assert self.window is not None
            # The following line copies our drawings from `canvas` to the visible window
            self.window.blit(canvas, canvas.get_rect())
            pygame.event.pump()
            pygame.display.update()

            # We need to ensure that human-rendering occurs at the predefined framerate.
            # The following line will automatically add a delay to keep the framerate stable.
            self.clock.tick(self.metadata["render_fps"])
        else:  # rgb_array
            return np.transpose(
                np.array(pygame.surfarray.pixels3d(canvas)), axes=(1, 0, 2)
            )

Close#

The close method should close any open resources that were used by the environment. In many cases, you don’t actually have to bother to implement this method. However, in our example render_mode may be "human" and we might need to close the window that has been opened:

    def close(self):
        if self.window is not None:
            import pygame 
            
            pygame.display.quit()
            pygame.quit()

In other environments close might also close files that were opened or release other resources. You shouldn’t interact with the environment after having called close.

Registering Envs#

In order for the custom environments to be detected by Gym, they must be registered as follows. We will choose to put this code in gym-examples/gym_examples/__init__.py.

from gym.envs.registration import register

register(
    id='gym_examples/GridWorld-v0',
    entry_point='gym_examples.envs:GridWorldEnv',
    max_episode_steps=300,
)

The environment ID consists of three components, two of which are optional: an optional namespace (here: gym_examples), a mandatory name (here: GridWorld) and an optional but recommended version (here: v0). It might have also been registered as GridWorld-v0 (the recommended approach), GridWorld or gym_examples/GridWorld, and the appropriate ID should then be used during environment creation.

The keyword argument max_episode_steps=300 will ensure that GridWorld environments that are instantiated via gym.make will be wrapped in a TimeLimit wrapper (see the wrapper documentation for more information). A done signal will then be produced if the agent has reached the target or 300 steps have been executed in the current episode. To distinguish truncation and termination, you can check info["TimeLimit.truncated"].

Apart from id and entrypoint, you may pass the following additional keyword arguments to register:

Name

Type

Default

Description

reward_threshold

float

None

The reward threshold before the task is considered solved

nondeterministic

bool

False

Whether this environment is non-deterministic even after seeding

max_episode_steps

int

None

The maximum number of steps that an episode can consist of. If not None, a TimeLimit wrapper is added

order_enforce

bool

True

Whether to wrap the environment in an OrderEnforcing wrapper

autoreset

bool

False

Whether to wrap the environment in an AutoResetWrapper

kwargs

dict

{}

The default kwargs to pass to the environment class

Most of these keywords (except for max_episode_steps, order_enforce and kwargs) do not alter the behavior of environment instances but merely provide some extra information about your environment. After registration, our custom GridWorldEnv environment can be created with env = gym.make('gym_examples/GridWorld-v0').

gym-examples/gym_examples/envs/__init__.py should have:

from gym_examples.envs.grid_world import GridWorldEnv

Creating a Package#

The last step is to structure our code as a Python package. This involves configuring gym-examples/setup.py. A minimal example of how to do so is as follows:

from setuptools import setup

setup(name='gym_examples',
    version='0.0.1',
    install_requires=['gym==0.23.1', 'pygame==2.1.0']
)

Creating Environment Instances#

After you have installed your package locally with pip install -e gym-examples, you can create an instance of the environment via:

import gym_examples
env = gym.make('gym_examples/GridWorld-v0')

You can also pass keyword arguments of your environment’s constructor to gym.make to customize the environment. In our case, we could do:

env = gym.make('gym_examples/GridWorld-v0', size=10)

Sometimes, you may find it more convenient to skip registration and call the environment’s constructor yourself. Some may find this approach more pythonic and environments that are instantiated like this are also perfectly fine (but remember to add wrappers as well!).

Using Wrappers#

Oftentimes, we want to use different variants of a custom environment, or we want to modify the behavior of an environment that is provided by Gym or some other party. Wrappers allow us to do this without changing the environment implementation or adding any boilerplate code. Check out the wrapper documentation for details on how to use wrappers and instructions for implementing your own. In our example, observations cannot be used directly in learning code because they are dictionaries. However, we don’t actually need to touch our environment implementation to fix this! We can simply add a wrapper on top of environment instances to flatten observations into a single array:

import gym_examples
from gym.wrappers import FlattenObservation

env = gym.make('gym_examples/GridWorld-v0')
wrapped_env = FlattenObservation(env)
print(wrapped_env.reset())     # E.g.  [3 0 3 3]

Wrappers have the big advantage that they make environments highly modular. For instance, instead of flattening the observations from GridWorld, you might only want to look at the relative position of the target and the agent. In the section on ObservationWrappers we have implemented a wrapper that does this job. This wrapper is also available in gym-examples:

import gym_examples
from gym_examples.wrappers import RelativePosition

env = gym.make('gym_examples/GridWorld-v0')
wrapped_env = RelativePosition(env)
print(wrapped_env.reset())     # E.g.  [-3  3]